Interpolation-Based Conditioning of Flow Matching Models for Bioisosteric Ligand Design
Abstract
Fast, unconditional 3D generative models can now produce high-quality molecules, but adapting them for specific design tasks often requires costly retraining. To address this, we introduce Interpolate-Integrate and Replacement Guidance, two training-free, inference-time conditioning strategies that provide control over E(3)-equivariant flow-matching models. Our methods generate bioisosteric 3D molecules by conditioning on seed ligands or fragment sets to preserve key determinants like shape and pharmacophore patterns, without requiring the original fragment atoms to be present. We demonstrate their effectiveness on three drug-relevant tasks: natural product ligand hopping, bioisosteric fragment merging, and pharmacophore merging.
Cite
Text
Ziv et al. "Interpolation-Based Conditioning of Flow Matching Models for Bioisosteric Ligand Design." International Conference on Learning Representations, 2026.Markdown
[Ziv et al. "Interpolation-Based Conditioning of Flow Matching Models for Bioisosteric Ligand Design." International Conference on Learning Representations, 2026.](https://mlanthology.org/iclr/2026/ziv2026iclr-interpolationbased/)BibTeX
@inproceedings{ziv2026iclr-interpolationbased,
title = {{Interpolation-Based Conditioning of Flow Matching Models for Bioisosteric Ligand Design}},
author = {Ziv, Yael and Buttenschoen, Martin and Scheibelberger, Lukas and Marsden, Brian and Deane, Charlotte},
booktitle = {International Conference on Learning Representations},
year = {2026},
url = {https://mlanthology.org/iclr/2026/ziv2026iclr-interpolationbased/}
}